Nanopore sequencing

Nanopore sequencing is a method of reading the order of bases in DNA or RNA by passing a single strand through a pore a nanometer wide, in a membrane, and recording the electric current. The bases in the narrowest part of the pore change the current, each base or short run of bases leaving a characteristic level, and software turns the trace into a sequence. In the best-known instruments a motor protein feeds the strand through a protein pore at about 400 bases a second and a neural network decodes the signal.^[1]

The method reads native DNA and RNA without amplification or fluorescent labels, gives reads as long as the molecules that survive preparation, and returns data while a run is under way. Its raw reads were at first far less accurate than those of short-read sequencing, and their maker now reports modal accuracies above 99%. Oxford Nanopore Technologies has supplied instruments since the MinION reached its first users in 2014, and other companies have since launched nanopore sequencers of their own. No solid-state pore has yet been reported to read an unknown DNA sequence base by base. A review of 2020 put it that the signal-to-noise ratio "has not yet allowed de novo DNA sequencing with solid-state pores", and a review of 2025 lists limited spatial resolution, fast translocation and noise as the obstacles.^[2][3] Molecular Reality Corporation builds solid-state pores and says of the goal in its guide: "That is the plan, and I flag it as a plan."^[4]

The idea of reading a strand through a pore was first proposed in the 1980s.^[5] In 1996 single strands of DNA and RNA were detected crossing a protein channel, and the authors concluded: "With further improvements, the method could in principle provide direct, high-speed detection of the sequence of bases in single molecules of DNA or RNA."^[6] Enzymes that slowed the strand arrived in 2012, the MinION reached its first users in 2014, and in 2016 the company changed to the R9 pore, derived from the bacterial protein CsgG.^[7] The full account is in History of nanopore sequencing. The first patent for nanopore sequencing, in the account of Deamer, Akeson and Branton, is US 5,795,782 of 1998, which names five inventors. One of them, George Church, is an investor in Molecular Reality Corporation.^[8][9][10]

A flow cell holds an array of membranes, each pierced by one protein nanopore and read through its own channel of an application-specific integrated circuit. Oxford Nanopore describes the membrane as "an electrically-resistant polymer membrane".^[11] A description of the MinION in 2016 gave its flow cell 2,048 pores in separate membranes, of which 512 are monitored at a time.^[9] A voltage across each membrane drives an ionic current through its pore. During library preparation, sequencing adapters pre-loaded with a motor protein are attached to the molecules to be read. The motor protein associates with a pore and controls the speed at which the molecule passes through it.^[12] With double-stranded DNA the motor unzips the duplex and passes one strand through the pore.^[5]

The motor is the difference between detecting DNA and reading it. A strand driven through a pore by the field alone moves each base past the sensing region in less than about 1 to 10 µs, and identifying a base from currents of picoamperes takes more than 100 µs.^[9] An enzyme at the mouth of the pore, a helicase or a polymerase, ratchets the strand through in discrete steps. The first combinations of an enzyme with a protein pore that read a strand step by step were reported in 2012.^[13][14] At 400 bases a second, the speed of current kits, each base stays in the pore for 2.5 ms, hundreds of times longer than in unassisted passage, and the data are sampled at 5 kHz, which gives 12.5 samples a base.^[1]

Basecalling turns the raw signal, called the squiggle, into letters. The first basecallers cut the trace into events and applied hidden Markov models, and later ones use recurrent and convolutional neural networks.^[5] The present basecaller, Dorado, comes in three sizes of model, fast, hac and sup, of which sup is the most accurate and the slowest. A model's name records the chemistry and speed it was trained for, as in dna_r10.4.1_e8.2_400bps_sup, for the R10.4.1 pore with Kit 14 at 400 bases a second, followed by a version number.^[1] A comparison of neural-network basecallers in 2019 found that training on data from one taxon raised consensus accuracy mostly by removing errors in methylation motifs, and that a larger network improved accuracy at a cost in speed.^[18]

Because the strand itself is in the pore, chemical modifications that change the current can be called from the same data. Dorado carries models for 5-methylcytosine and 5-hydroxymethylcytosine, 4-methylcytosine and 6-methyladenine.^[1] Calling 5-methylcytosine from MinION data was reported in 2017 by two groups in the same issue of Nature Methods.^[19][20] RNA can be read directly, without conversion to DNA.^[21]

Accuracy is quoted as a percentage of correct bases or as a Phred quality score Q, where Q20 means one error in 100 bases and Q30 one in 1,000. Raw reads have improved with each generation of pore, motor and basecaller.

The 2021 duplex figure refers to reads in which the two strands of one molecule are read in turn and combined, and the maker reported that the longest read at Q30 so far was 156 kilobases.^[23] Independent tests of the Q20+ chemistry were more modest at first. On amplicons of SARS-CoV-2, the average accuracy rose from 96.52% for the earlier pore and kit to 98.34%, and the share of reads above 99% from 0.61% to 30.1%, with the consensus matching short-read sequencing.^[24] A test on bacterial genomes with the Kit 14 chemistry found that assemblies from nanopore reads alone were accurate at a depth of 40-fold or more, using updated basecallers that had been trained on methylated bacterial DNA to fix common systematic substitutions.^[25] The maker's accuracy page states that raw reads "now achieve 99.75% (Q26) accuracy" with the latest models.^[26]

Nothing in the method fixes a read length. The maker says that a read "is only limited by the length of the DNA or RNA presented for sequencing", gives the range as 20 to 4,000,000 bases, and says samples can be prepared in as little as 10 minutes. Because the strand is not amplified, native DNA and RNA are analyzed directly and keep their base modifications.^[12] Read length and portability opened the first applications. In April 2015 a sequencing laboratory transported in standard airline luggage to Guinea sequenced 142 samples of Ebola virus, returning results less than 24 hours after a positive sample arrived, with sequencing taking as little as 15 to 60 minutes, for genomic surveillance (Biosurveillance).^[27] A MinION also went to the International Space Station, where nine runs over six months produced 276,882 reads, and the genome of Escherichia coli assembled from the raw reads into a single contig covering 99.9% of the genome at 98.6% identity.^[28] Ultra-long reads were used to finish the human X chromosome, including its centromeric satellite array of about 3.1 Mb, the first human chromosome assembled from end to end.^[29] Because the pore reads a strand while it moves, software can also choose what to read, ejecting a molecule it does not want by reversing the voltage.^[30]

Oxford Nanopore Technologies began in 2005 as a spinout from the University of Oxford's Department of Chemistry and floated on the London Stock Exchange in 2021.^[31] Its instruments are the MinION, the GridION and the PromethION, and its smallest flow cell is the Flongle. A 2024 study of the QNome sequencer of Qitan Technology, a company in Chengdu, described it as making Qitan the second supplier of nanopore sequencers.^[32] Axbio, of Santa Clara, California, lists its AxiLona AXP-100 as a nanopore gene sequencer and describes a Bio-CMOS sequencing chip.^[33] On June 29, 2026 Roche launched AXELIOS 1, for research use only, built on its Sequencing by expansion chemistry. That chemistry converts DNA or RNA into surrogate molecules about fifty times longer, which a CMOS sensor with millions of nanopores reads, and Roche gives reads of up to about 1,500 bases.^[34] The leading methods that do not use nanopores are the sequencing by synthesis of Illumina and the single-molecule real-time sequencing of Pacific Biosciences.

A flow cell is a consumable with a clock on it. Oxford Nanopore says to store flow cells unopened at 2 to 8 °C, gives each a warranty of 12 weeks from receipt, and says a flow cell holds buffer to run for about 72 hours under optimal conditions; it replaces an unused flow cell that has fewer than 800 nanopores at the start.^[17] In a 2021 study of what was then the mature chemistry, homopolymers and short repeats were the source of about half of all errors, which were mostly deletions.^[22] Larger basecalling models take more time, which is why Dorado offers three sizes.^[1]

Pores in silicon nitride and other inorganic membranes survive heat, dryness and chemistry that destroy protein pores, and their size can be chosen at fabrication. They are also noisier and less uniform, and DNA crosses them too fast to read.^[2] In ultrathin membranes of graphene or molybdenum disulfide, DNA moves at 3 to 50 bases a microsecond, and reading at that speed needs megahertz bandwidths.^[35] An amplifier built for the purpose has recorded at 10 MHz, with a signal-to-noise ratio above 10 for single-stranded DNA at 5 MHz in sub-2 nm pores, but a base spends only a few hundred nanoseconds in the pore.^[36] The length of the sensing region is a second obstacle. In a solid-state pore it is the thickness of the membrane: about 5 nm for silicon nitride, against 0.7 nm for molybdenum disulfide.^[2] A stem 5 nm long, as in alpha-hemolysin, puts about 12 bases into each current level, and silicon nitride pores are about as long.^[9]

Each step of the problem has had partial results. Silicon nitride pores 0.8 to 2 nm wide told apart 30-base homopolymers of A, C and T at a signal-to-noise ratio above 10 at 1 MHz, a result the authors called "an important milestone" while noting that identifying sequences within a molecule would need further improvements in sensitivity and noise.^[37] A viscosity gradient of ionic liquid slowed DNA through a molybdenum disulfide pore enough to identify all four nucleotides statistically, with a signal-to-noise ratio above 10.^[35] Osmium tags on pyrimidine bases gave strands different blockade amplitudes in a silicon nitride pore according to how many bases were tagged, which the authors presented as a proof of principle.^[38] Genome mapping finds the positions of tags along long DNA without reading the bases between them. With pairs of solid-state pores a company-led team, working with a university laboratory, reached 3.1-fold coverage across 2,355 resolvable sites of the 4.6 Mb E. coli chromosome in 5.6 hours of recording, with an intertag resolution of 150 bp in engineered DNA.^[39] What remains unpublished is a base-by-base read of an unknown sequence.

Molecular Reality Corporation's approach is to slow the molecule mechanically. A juxtapore is "a pore formed by two surfaces sliding alongside each other", and the company's guide says of true sequencing: "No solid state pore has done it commercially yet. Our mechanical valve work on juxtapores exists precisely because reading requires slowing the molecule down, and slowing molecules down is what sliding occlusion is for."^[4] A feature in Signal & Noise of October 2026 concluded that holding a molecule still long enough to read it is the step between sliding occlusion and sequencing, and that no one has shown it with a juxtapore.^[40]

• Biological nanopore
• Motor protein (nanopore)
• Protein sequencing
• Nanopore protein sequencing
• Machine learning on nanopore signals
• List of nanopore companies

• Branton, D.; et al. (2008). "The potential and challenges of nanopore sequencing". Nature Biotechnology 26 (10): 1146-1153. doi:10.1038/nbt.1495.
• Jain, M.; Olsen, H. E.; Paten, B.; Akeson, M. (2016). "The Oxford Nanopore MinION: delivery of nanopore sequencing to the genomics community". Genome Biology 17: 239. doi:10.1186/s13059-016-1103-0.